Abstract
Face recognition remains a significant challenge in computer vision due to high-dimensional data and limited training samples, often referred to as the "Small Sample Size" (SSS) problem. This paper explores subspace methods to address three critical issues: matrix singularity in Linear Discriminant Analysis (LDA), over-fitting of discriminant vectors, and sensitivity to image occlusions. We propose a multi-faceted approach involving Bi-Directional Principal Component Analysis (BDPCA) combined with LDA, a post-processing Gaussian filter for regularization, and Iteratively Reweighted Fitting (IRF) for robustness against noise and occlusions.
Introduction
Holistic-based face recognition extracts features from the entire face region, offering robustness against noise and blurring. However, the high dimensionality of facial images leads to the "curse of dimensionality." Subspace methods aim to reduce this dimensionality while preserving discriminative information. In practice, face recognition systems often face the "Small Sample Size" (SSS) problem, where the number of available training images per person is much smaller than the dimensionality of the feature space. This leads to mathematical instabilities and poor generalization.
Challenges in Subspace-Based Recognition
Traditional subspace methods like Linear Discriminant Analysis (LDA) rely on the analysis of scatter matrices to find optimal projection vectors.
The Singularity Problem
LDA seeks to maximize the ratio of between-class scatter ($S_b$) to within-class scatter ($S_w$). In SSS scenarios, $S_w$ often becomes singular (non-invertible) because the number of samples is insufficient to span the high-dimensional space. This prevents the direct calculation of the optimal discriminant vectors, causing standard mathematical solvers to fail.
Over-fitting and Noise Sensitivity
Discriminant vectors derived from limited data can become "noisy" or "wiggly," capturing high-frequency pixel variations that do not represent actual facial features. Furthermore, standard methods like Eigenfaces are highly sensitive to partial occlusions (e.g., sunglasses, scarves) and non-Gaussian noise, as they treat all pixels with equal importance, regardless of whether they are part of the face or an obstruction.
Proposed Methodology
To overcome these limitations, we introduce three key enhancements to the subspace framework.
BDPCA+LDA for Efficient Feature Extraction
Instead of flattening an $80 \times 80$ image into a 6,400-dimensional vector, Bi-Directional PCA (BDPCA) operates directly on image matrices. By projecting the image matrix onto two lower-dimensional subspaces simultaneously, BDPCA preserves spatial structure and significantly reduces the dimensionality before LDA is applied. This approach mitigates the singularity problem and reduces computational costs.
Post-processed LDA (PEFM)
To regularize the learning process, we apply a 2D-Gaussian filter to the discriminant vectors. This "smoothing" operation incorporates a spatial prior, ensuring that neighboring pixels contribute coherently to the classification. This method, termed the Post-processed Eigenface Method (PEFM), effectively reduces over-fitting and improves recognition accuracy by focusing on broader facial structures rather than isolated pixel noise. This is theoretically linked to the Image Euclidean Distance (IMED) metric.
Robustness via IRF-Eigenfaces
For handling occlusions, we replace standard linear projection with Iteratively Reweighted Fitting (IRF). Using an Iteratively Reweighted Least Squares (IRLS) algorithm, the model identifies occluded regions as statistical outliers and assigns them lower weights. This allows for accurate recognition even when significant portions of the face are hidden by accessories like sunglasses or scarves.
Experimental Results
The proposed methods were evaluated on the FERET and AR face databases.
- Efficiency & Accuracy: On the FERET dataset, BDPCA+LDA achieved an Average Recognition Rate (ARR) of 87.14%, outperforming standard PCA+LDA (84.69%). Training time was slashed from 254.2 seconds to just 57.5 seconds.
- Extreme Noise Robustness: In tests with 50% "salt and pepper" noise, standard Eigenfaces' accuracy plummeted to 36.67%. IRF-Eigenfaces maintained a staggering 95.83% accuracy.
- Handling Occlusion: On the AR database, IRF-Eigenfaces achieved over 91% accuracy for faces wearing scarves, compared to only 35% for standard Eigenfaces.
Conclusion
By addressing singularity through matrix-based dimensionality reduction, over-fitting through spatial regularization, and occlusion through robust iterative fitting, we significantly enhance the performance of subspace methods for face recognition. Future work will focus on integrating these techniques into a unified framework capable of handling simultaneous variations in pose, illumination, and expression.
Summary updated